CSE 255 Lecture 5 Data Mining and Predictive Analytics Recommender - - PowerPoint PPT Presentation

CSE 255 – Lecture 5

Data Mining and Predictive Analytics

Recommender Systems

Why recommendation? The goal of recommender systems is…

• To help people discover new content

Why recommendation? The goal of recommender systems is…

• To help us find the content we were

already looking for

Are these recommendations good or bad?

Why recommendation? The goal of recommender systems is…

• To discover which things go together

Why recommendation? The goal of recommender systems is…

• To personalize user experiences in

response to user feedback

Why recommendation? The goal of recommender systems is…

• To recommend incredible products

that are relevant to our interests

Why recommendation? The goal of recommender systems is…

• To identify things that we like

Why recommendation? The goal of recommender systems is…

• To help people discover new content
• To help us find the content we were

already looking for

• To discover which things go together
• To personalize user experiences in

response to user feedback

• To identify things that we like

To model people’s preferences, opinions, and behavior

Recommending things to people Suppose we want to build a movie recommender

e.g. which of these films will I rate highest?

Recommending things to people We already have a few tools in our “supervised learning” toolbox that may help us

Recommending things to people

Movie features: genre, actors, rating, length, etc. User features: age, gender, location, etc.

Recommending things to people With the models we’ve seen so far, we can build predictors that account for…

• Do women give higher ratings than men?
• Do Americans give higher ratings than Australians?
• Do people give higher ratings to action movies?
• Are ratings higher in the summer or winter?
• Do people give high ratings to movies with Vin Diesel?

So what can’t we do yet?

Recommending things to people Consider the following linear predictor (e.g. from week 1):

Recommending things to people But this is essentially just two separate predictors!

user predictor movie predictor

That is, we’re treating user and movie features as though they’re independent

Recommending things to people But these predictors should (obviously?) not be independent

do I tend to give high ratings? does the population tend to give high ratings to this genre of movie?

But what about a feature like “do I give high ratings to this genre of movie”?

Recommending things to people

Recommender Systems go beyond the methods we’ve seen so far by trying to model the relationships between people and the items they’re evaluating my (user’s) “preferences” HP’s (item) “properties”

preference Toward “action” preference toward “special effects” is the movie action- heavy? are the special effects good? Compatibility

T

• day

Recommender Systems 1. Collaborative filtering

(performs recommendation in terms of user/user and item/item similarity)

2. Latent-factor models

(performs recommendation by projecting users and items into some low-dimensional space)

Defining similarity between users & items Q: How can we measure the similarity between two users? A: In terms of the items they purchased! Q: How can we measure the similarity between two items? A: In terms of the users who purchased them!

Defining similarity between users & items e.g.: Amazon

Definitions Definitions

= set of items purchased by user u = set of users who purchased item i

Definitions

Or equivalently… users items = binary representation items purchased by u = binary representation of users who purchased i

• 0. Euclidean distance

A B

Euclidean distance:

e.g. between two items i,j (similarly defined between two users)

• 0. Euclidean distance

Euclidean distance:

e.g.: U_1 = {1,4,8,9,11,23,25,34} U_2 = {1,4,6,8,9,11,23,25,34,35,38} U_3 = {4} U_4 = {5} Problem: favors small sets, even if they have few elements in common

• 1. Jaccard similarity

A B  Maximum of 1 if the two users purchased exactly the same set of items

(or if two items were purchased by the same set of users)

 Minimum of 0 if the two users purchased completely disjoint sets of items

(or if the two items were purchased by completely disjoint sets of users)

• 2. Cosine similarity

(vector representation of users who purchased harry potter)

(theta = 0)  A and B point in exactly the same direction (theta = 180)  A and B point in opposite directions (won’t actually happen for 0/1 vectors) (theta = 90)  A and B are

• rthogonal

• 2. Cosine similarity

Why cosine?

• Unlike Jaccard, works for arbitrary vectors
• E.g. what if we have opinions in addition to purchases?

bought and liked didn’t buy bought and hated

• 2. Cosine similarity

(vector representation of users’ ratings of Harry Potter)

(theta = 0)  Rated by the same users, and they all agree (theta = 180)  Rated by the same users, but they completely disagree about it (theta = 90)  Rated by different sets of users

E.g. our previous example, now with “thumbs-up/thumbs-down” ratings

• 4. Pearson correlation

What if we have numerical ratings (rather than just thumbs-up/down)?

bought and liked didn’t buy bought and hated

• 4. Pearson correlation

What if we have numerical ratings (rather than just thumbs-up/down)?

• We wouldn’t want 1-star ratings to be parallel to 5-

star ratings

• So we can subtract the average – values are then

negative for below-average ratings and positive for above-average ratings

items rated by both users average rating by user v

• 4. Pearson correlation

Compare to the cosine similarity:

Pearson similarity (between users): Cosine similarity (between users):

items rated by both users average rating by user v

Collaborative filtering in practice

How did Amazon generate their ground-truth data?

Given a product: Let be the set of users who viewed it

Rank products according to: (or cosine/pearson)

.86 .84 .82 .79 … Linden, Smith, & York (2003)

Collaborative filtering in practice Note: (surprisingly) that we built something pretty useful out of nothing but rating data – we didn’t look at any features of the products whatsoever

Collaborative filtering in practice But: we still have a few problems left to address…

1. This is actually kind of slow given a huge enough dataset – if one user purchases one item, this will change the rankings of every

• ther item that was purchased by at least
• ne user in common

2. Of no use for new users and new items (“cold- start” problems 3. Won’t necessarily encourage diverse results

Questions

CSE 255 – Lecture 5

Data Mining and Predictive Analytics

Latent-factor models

Latent factor models So far we’ve looked at approaches that try to define some definition of user/user and item/item similarity Recommendation then consists of

• Finding an item i that a user likes (gives a high rating)
• Recommending items that are similar to it (i.e., items j

with a similar rating profile to i)

Latent factor models What we’ve seen so far are unsupervised approaches and whether the work depends highly on whether we chose a “good” notion of similarity So, can we perform recommendations via supervised learning?

Latent factor models e.g. if we can model Then recommendation will consist of identifying

The Netflix prize

In 2006, Netflix created a dataset of 100,000,000 movie ratings Data looked like: The goal was to reduce the (R)MSE at predicting ratings: Whoever first manages to reduce the RMSE by 10% versus Netflix’s solution wins $1,000,000

model’s prediction ground-truth

This led to a lot of research on rating prediction by minimizing the Mean- Squared Error

(it also led to a lawsuit against Netflix, once somebody managed to de-anonymize their data)

We’ll look at a few of the main approaches The Netflix prize

Rating prediction Let’s start with the simplest possible model:

user item

Here the RMSE is just equal to the standard deviation of the data

(and we cannot do any better with a 0th order predictor)

Rating prediction What about the 2nd simplest model?

user item how much does this user tend to rate things above the mean? does this item tend to receive higher ratings than others

e.g.

Rating prediction The optimization problem becomes: Jointly convex in \beta_i, \beta_u. Can be solved by iteratively removing the mean and solving for beta

error regularizer

Rating prediction Iterative procedure – repeat the following updates until convergence:

(exercise: write down derivatives and convince yourself of these update equations!)

Rating prediction

user predictor movie predictor

Looks good (and actually works surprisingly well), but doesn’t solve the basic issue that we started with That is, we’re still fitting a function that treats users and items independently

Recommending things to people How about an approach based on dimensionality reduction?

my (user’s) “preferences” HP’s (item) “properties” i.e., let’s come up with low-dimensional representations of the users and the items so as to best explain the data

Dimensionality reduction We already have some tools that ought to help us, e.g. from lecture 3:

What is the best low- rank approximation of R in terms of the mean- squared error?

Dimensionality reduction We already have some tools that ought to help us, e.g. from lecture 3:

eigenvectors of eigenvectors of (square roots of) eigenvalues of

Singular Value Decomposition The “best” rank-K approximation (in terms of the MSE) consists

• f taking the eigenvectors with the highest eigenvalues

Dimensionality reduction But! Our matrix of ratings is only partially

• bserved; and it’s really big!

Missing ratings

SVD is not defined for partially observed matrices, and it is not practical for matrices with 1Mx1M+ dimensions

; and it’s really big!

Latent-factor models Instead, let’s solve approximately using gradient descent

items users

K-dimensional representation

• f each user

K-dimensional representation

• f each item

Latent-factor models

my (user’s) “preferences” HP’s (item) “properties”

Let’s write this as:

Latent-factor models Let’s write this as: Our optimization problem is then Problem: this is certainly not convex

(proof is easy: (1) it is smooth; (2) permuting the columns of gamma preserves the objective; (3) therefore it has multiple local optima and cannot be convex; (4) in other words it must look like this: ) error regularizer

permutations of local minima

Latent-factor models Oh well. We’ll just solve it approximately Observation: if we know either the user

• r the item parameters, the problem

becomes easy

e.g. fix gamma_i – pretend we’re fitting parameters for features

Latent-factor models This gives rise to a simple (though approximate) solution

1) fix . Solve 2) fix . Solve 3,4,5…) repeat until convergence

• bjective:

Each of these subproblems is “easy” – just regularized least-squares, like we’ve been doing since week 1. This procedure is called alternating least squares.

Latent-factor models

Movie features: genre, actors, rating, length, etc. User features: age, gender, location, etc.

Observation: we went from a method which uses only features: to one which completely ignores them:

Latent-factor models Should we use features or not? 1) Argument against features:

Imagine incorporating features into the model like:

which is equivalent to: knowns unknowns but this has fewer degrees of freedom than a model which replaces the knowns by unknowns:

Latent-factor models Should we use features or not? 1) Argument against features:

So, the addition of features adds no expressive power to the

• model. We could have a feature like “is this an action

movie?”, but if this feature were useful, the model would “discover” a latent dimension corresponding to action movies, and we wouldn’t need the feature anyway In the limit, this argument is valid: as we add more ratings per user, and more ratings per item, the latent-factor model should automatically discover any useful dimensions of variation, so the influence of observed features will disappear

Latent-factor models Should we use features or not? 2) Argument for features:

But! Sometimes we don’t have many ratings per user/item Latent-factor models are next-to-useless if either the user or the item was never observed before

reverts to zero if we’ve never seen the user before (because of the regularizer)

Latent-factor models Should we use features or not? 2) Argument for features:

This is known as the cold-start problem in recommender

• systems. Features are not useful if we have many
• bservations about users/items, but are useful for new users

and items. We also need some way to handle users who are active, but don’t necessarily rate anything, e.g. through implicit feedback

Overview & recap Tonight we’ve followed the programme below:

• 1. Measuring similarity between users/items for

binary prediction (e.g. Jaccard similarity)

• 2. Measuring similarity between users/items for real-

valued prediction (e.g. cosine/Pearson similarity)

• 3. Dimensionality reduction for real-valued

prediction (latent-factor models)

• 4. Finally – dimensionality reduction for binary

prediction

One-class recommendation How can we use dimensionality reduction to predict binary

• utcomes?
• In weeks 1&2 we saw regression and logistic
• regression. These two approaches use the same

type of linear function to predict real-valued and binary outputs

• We can apply an analogous approach to binary

recommendation tasks

One-class recommendation This is referred to as “one-class” recommendation

• In weeks 1&2 we saw regression and logistic
• regression. These two approaches use the same

type of linear function to predict real-valued and binary outputs

• We can apply an analogous approach to binary

recommendation tasks

One-class recommendation Suppose we have binary (0/1) observations (e.g. purchases) or positive/negative feedback (thumbs-up/down)

• r

purchased didn’t purchase liked didn’t evaluate didn’t like

One-class recommendation So far, we’ve been fitting functions of the form

• Let’s change this so that we maximize the difference in

predictions between positive and negative items

• E.g. for a user who likes an item i and dislikes an item j we

want to maximize:

One-class recommendation We can think of this as maximizing the probability of correctly predicting pairwise preferences, i.e.,

• As with logistic regression, we can now maximize the

likelihood associated with such a model by gradient ascent

• In practice it isn’t feasible to consider all pairs of

positive/negative items, so we proceed by stochastic gradient ascent – i.e., randomly sample a (positive, negative) pair and update the model according to the gradient w.r.t. that pair

Summary Recap

• 1. Measuring similarity between users/items for

binary prediction Jaccard similarity

• 2. Measuring similarity between users/items for real-

valued prediction cosine/Pearson similarity

• 3. Dimensionality reduction for real-valued prediction

latent-factor models

• 4. Dimensionality reduction for binary prediction
• ne-class recommender systems

Questions? Further reading:

One-class recommendation: http://goo.gl/08Rh59 Amazon’s solution to collaborative filtering at scale: http://www.cs.umd.edu/~samir/498/Amazon-Recommendations.pdf

An (expensive) textbook about recommender systems: http://www.springer.com/computer/ai/book/978-0-387-85819-7 Cold-start recommendation (e.g.): http://wanlab.poly.edu/recsys12/recsys/p115.pdf

CSE 255 – Lecture 5

Data Mining and Predictive Analytics

Extensions of latent-factor models, (and more on the Netflix prize!)

Extensions of latent-factor models So far we have a model that looks like: How might we extend this to:

• Incorporate features about users and items
• Handle implicit feedback
• Change over time

See Yehuda Koren (+Bell & Volinsky)’s magazine article: “Matrix Factorization Techniques for Recommender Systems” IEEE Computer, 2009

Extensions of latent-factor models 1) Features about users and/or items

(simplest case) Suppose we have binary attributes to describe users or items

A(u) = [1,0,1,1,0,0,0,0,0,1,0,1]

attribute vector for user u e.g. is female is male is between 18-24yo

Extensions of latent-factor models 1) Features about users and/or items

(simplest case) Suppose we have binary attributes to describe users or items

• Associate a parameter vector with each attribute
• Each vector encodes how much a particular feature

“offsets” the given latent dimensions

A(u) = [1,0,1,1,0,0,0,0,0,1,0,1]

attribute vector for user u e.g. y_0 = [-0.2,0.3,0.1,-0.4,0.8] ~ “how does being male impact gamma_u”

Extensions of latent-factor models 1) Features about users and/or items

(simplest case) Suppose we have binary attributes to describe users or items

• Associate a parameter vector with each attribute
• Each vector encodes how much a particular feature

“offsets” the given latent dimensions

• Model looks like:
• Fit as usual:

error regularizer

Extensions of latent-factor models 2) Implicit feedback

Perhaps many users will never actually rate things, but may still interact with the system, e.g. through the movies they view, or the products they purchase (but never rate)

• Adopt a similar approach – introduce a binary vector

describing a user’s actions

N(u) = [1,0,0,0,1,0,….,0,1]

implicit feedback vector for user u e.g. y_0 = [-0.1,0.2,0.3,-0.1,0.5] Clicked on “Love Actually” but didn’t watch

Extensions of latent-factor models 2) Implicit feedback

Perhaps many users will never actually rate things, but may still interact with the system, e.g. through the movies they view, or the products they purchase (but never rate)

• Adopt a similar approach – introduce a binary vector

describing a user’s actions

• Model looks like:

normalize by the number of actions the user performed

Extensions of latent-factor models 3) Change over time

There are a number of reasons why rating data might be subject to temporal effects…

Extensions of latent-factor models 3) Change over time

Netflix ratings

• ver time

early 2004

Figure from Koren: “Collaborative Filtering with Temporal Dynamics” (KDD 2009)

Netflix changed their interface!

Extensions of latent-factor models 3) Change over time

Netflix ratings by movie age

Figure from Koren: “Collaborative Filtering with Temporal Dynamics” (KDD 2009)

People tend to give higher ratings to older movies

Extensions of latent-factor models 3) Change over time

A few temporal effects from beer reviews

Extensions of latent-factor models 3) Change over time

There are a number of reasons why rating data might be subject to temporal effects…

e.g. “Collaborative filtering with temporal dynamics” Koren, 2009

• Changes in the interface
• People give higher ratings to older movies (or, people

who watch older movies are a biased sample)

• The community’s preferences gradually change over time
• My girlfriend starts using my Netflix account one day
• I binge watch all 144 episodes of buffy one week and

then revert to my normal behavior

• I become a “connoisseur” of a certain type of movie
• Anchoring, public perception, seasonal effects, etc.

e.g. “Sequential & temporal dynamics of online opinion” Godes & Silva, 2012 e.g. “Temporal recommendation on graphs via long- and short-term preference fusion” Xiang et al., 2010 e.g. “Modeling the evolution

• f user expertise through
• nline reviews”

McAuley & Leskovec, 2013

Extensions of latent-factor models 3) Change over time

Each definition of temporal evolution demands a slightly different model assumption (we’ll see some in more detail later tonight!) but the basic idea is the following: 1) Start with our original model: 2) And define some of the parameters as a function of time: 3) Add a regularizer to constrain the time-varying terms:

parameters should change smoothly (I’ll give an example in the set of slides after the break)

Extensions of latent-factor models 3) Change over time

After the break: how do people acquire tastes for beers (and potentially for other things) over time? Differences between “beginner” and “expert” preferences for different beer styles

Extensions of latent-factor models 4) Missing-not-at-random

• Our decision about whether to purchase a movie (or

item etc.) is a function of how we expect to rate it

• Even for items we’ve purchased, our decision to enter a

rating or write a review is a function of our rating

• e.g. some rating distribution from a few datasets:

EachMovie MovieLens Netflix

Figure from Marlin et al. “Collaborative Filtering and the Missing at Random Assumption” (UAI 2007)

Extensions of latent-factor models 4) Missing-not-at-random

e.g. Men’s watches:

Extensions of latent-factor models 4) Missing-not-at-random

• Our decision about whether to purchase a movie (or

item etc.) is a function of how we expect to rate it

• Even for items we’ve purchased, our decision to enter a

rating or write a review is a function of our rating

• So we can predict ratings more accurately by building

models that account for these differences

• 1. Not-purchased items have a different prior on ratings

than purchased ones

• 2. Purchased-but-not-rated items have a different prior on

ratings than rated ones

Figure from Marlin et al. “Collaborative Filtering and the Missing at Random Assumption” (UAI 2007)

Moral(s) of the story How much do these extension help?

bias terms implicit feedback temporal dynamics

Moral: increasing complexity helps a bit, but changing the model can help a lot

Figure from Koren: “Collaborative Filtering with Temporal Dynamics” (KDD 2009)

Moral(s) of the story So what actually happened with Netflix?

• The AT&T team “BellKor”, consisting of Yehuda Koren, Robert Bell, and Chris

Volinsky were early leaders. Their main insight was how to effectively incorporate temporal dynamics into recommendation on Netflix.

• Before long, it was clear that no one team would build the winning solution,

and Frankenstein efforts started to merge. Two frontrunners emerged, “BellKor’s Pragmatic Chaos”, and “The Ensemble”.

• The BellKor team was the first to achieve a 10% improvement in RMSE, putting

the competition in “last call” mode. The winner would be decided after 30 days.

• After 30 days, performance was evaluated on the hidden part of the test set.
• Both of the frontrunning teams had the same RMSE (up to some precision) but

BellKor’s team submitted their solution 20 minutes earlier and won $1,000,000 For a less rough summary, see the Wikipedia page about the Netflix prize, and the nytimes article about the competition: http://goo.gl/WNpy7o

Moral(s) of the story Afterword

• Netflix had a class-action lawsuit filed against them after somebody de-

anonymized the competition data

• $1,000,000 seems to be incredibly cheap for a company the size of Netflix in

terms of the amount of research that was devoted to the task, and the potential benefit to Netflix of having their recommendation algorithm improved by 10%

• Other similar competitions have emerged, such as the Heritage Health Prize

($3,000,000 to predict the length of future hospital visits)

• But… the winning solution never made it into production at Netflix – it’s a

monolithic algorithm that is very expensive to update as new data comes in*

*source: a friend of mine told me and I have no actual evidence of this claim

Moral(s) of the story Finally…

Q: Is the RMSE really the right approach? Will improving rating prediction by 10% actually improve the user experience by a significant amount? A: Not clear. Even a solution that only changes the RMSE slightly could drastically change which items are top-ranked and ultimately suggested to the user. Q: But… are the following recommendations actually any good? A1: Yes, these are my favorite movies!

• r A2: No! There’s no diversity, so how will I discover new content?

5.0 stars 5.0 stars 5.0 stars 5.0 stars 4.9 stars 4.9 stars 4.8 stars 4.8 stars

predicted rating

Summary Various extensions of latent factor models:

• Incorporating features

e.g. for cold-start recommendation

• Implicit feedback

e.g. when ratings aren’t available, but other actions are

• Incorporating temporal information into latent factor models

seasonal effects, short-term “bursts”, long-term trends, etc.

• Missing-not-at-random

incorporating priors about items that were not bought or rated

• The Netflix prize

Things I didn’t get to… Socially regularized recommender systems

see e.g. “Recommender Systems with Social Regularization”

http://research.microsoft.com/en-us/um/people/denzho/papers/rsr.pdf

social regularizer network

Things I didn’t get to…

Recommendation in certain settings (e.g. online advertising) has drastically different assumptions compared to what’s appropriate for products on Amazon or movies on Netflix. e.g.

• I want to show the ad I believe you are most likely to click on
• But, I also want to discover your preferences for categories of

ads about which I have no information

• So there is a natural exploration/exploitation tradeoff when

making recommendations See e.g. “A Contextual-Bandit Approach to Personalized News Article Recommendation”

http://www.research.rutgers.edu/~lihong/pub/Li10Contextual.pdf

Online advertising

Questions? Further reading:

Yehuda Koren’s, Robert Bell, and Chris Volinsky’s IEEE computer article: http://www2.research.att.com/~volinsky/papers/ieeecomputer.pdf Paper about the “Missing-at-Random” assumption, and how to address it: http://www.cs.toronto.edu/~marlin/research/papers/cfmar-uai2007.pdf Collaborative filtering with temporal dynamics: http://research.yahoo.com/files/kdd-fp074-koren.pdf Recommender systems and sales diversity: http://papers.ssrn.com/sol3/papers.cfm?abstract_id=955984

Up next: Assignment 2!